Beyond Traffic-Light Risk Scores: Building Climate Risk Models That Actually Work | Joan Saladich (Part 2)

If something is not insurable, no bank will finance this something if there is nobody ready to insure this, so it will be not developed. So if a hospital cannot be insured, have a big problem. I'm trying to think about what's important for the insurance industry or what I would ask because it's a very crucial industry for our society. Hello and welcome. My name is Jake Harding and you're listening to Making Risk Exploring the Ecosystem, a new companion series to the Making Risk Flow podcast where we sit down with data providers, technology leaders and cytora partners to uncover actionable knowledge on how insurance can achieve frictionless risk flows. So I guess onto the science then. How does a massive shift in underlying science like we've discussed, how does that expose the limitations of static climate indices? And what does that mean for carriers whose risk assessments are built around scenarios that science now deem impossible? So with the concept of static that you are asking, let's say there are two interpretations because that's kind of a double edged sword. On one hand in climate modeling, it's not weather forecasting. So climate modeling per se is static. That's the first interpretation. And I also see in some cases that there are different actors and colleagues in the industry that are trying to promote a more dynamic climate modeling. Like, hey, we are updating our climate models every year. In my humble view, this is also not the correct approach because imagine that you're modeling the next 100 years and you're updating based on the last 50 years. One year change shouldn't have a big impact on your model that runs for the next 100 years because your model is pretty much solid. Maybe a decade of data could have an impact on the trends, on the deviations, return periods, but just one single year, that could be an abnormal year. Yeah, an anomalous year. So per se, this is a concept that I believe it's very important because at some point it could mislead the end user, I guess not the AI, but the end user. It's like constantly saying that you are updating your model so that your model is not static in climate modeling and projections, not in weather forecasting or in seasonal forecasts or if somebody is doing. We tried a yearly forecast, but yearly predictions, but in the projections it shouldn't have an impact in your projection. Yeah, so it's static. Now the other idea is the silos idea. It's the other field due to this, as we mentioned at the beginning, traffic light factor or a nutritional score factor. Yeah, like what we see in supermarkets from 1 to 5. It seems like when you have a wildfire risk index, it's purely relying on on the weather or on the climate, on what's on the atmosphere. And there should be science on what's on the ground, which kind of vegetation, how will this vegetation evolve? So this concept of static, it's very interesting. On the other hand, let me put you a practical example, following wildfires, because as I was speaking, I just remember wildfire risk indexes. There are some risk indexes that were generated for weather forecasts and they are coupling atmosphere like temperature, rainfall, relative humidity, irradiance, blah, blah, blah. They are coupling this with soil data and for a weather forecast that's cool for the next five days. And you just have a relation. Yeah. So some point, soil data is almost assumed to be static. So the next five days, my vegetation, it's green today, it's going to be green in five days. But it happens a lot in climate projections that they just extrapolate the same index of wildfire risk with this same coupling. And this is a very wrong idea because vegetation growth and vegetation dies and it's more prone, let's say, to getting to fire when it dies. And this happens in Spain a lot. And wildfire risk indices per se are not working in Spain we have January and February these last two years. January and February were very, very rainy. And so if anybody was forecasting, let's say, the risk index, the wildfire risk index for June, for July, they were considering the status of the vegetation and soil humidity as January. But in July this will be dry as hell and all this vegetation will be dead. So the risk of wildfire increases a lot. And this is something that it's not even solved, I believe, from a scientific standpoint. Like how can we couple as we project the wildfire risk index with the vegetation? And this concept is also, I would also relate it to this static, like it has to be coupled because if not, you can say, oh, there is no risk, it was from a weather standpoint or from a seasonal standpoint, seasonal forecast there was no risk, or there was lower risk. But actually all your vegetation, it was raining a lot, the vegetation grow a lot, and then by that point it was very susceptible or prone to transform into fire. So this is something that it's not being considered. So this concept of static, at some point, it's something that we have to work out. But in terms of static, of reinitializing or updating the climate change model, we have to be careful here too. Those are two different statics I don't know if I explained myself properly because it was a little bit complex. Totally understand. And it really highlights. We've spoken about AI's ability to integrate a wide variety of different data points. But then the next dimension to that is those data points exist over time. And so there's a second dimension which exponentially increases the amount of input that you need to process to come up with kind of reliable output. And then I think a third point that you touched on is asking a human to do that. We are prone to recency bias. As you say, like one year's worth of slightly different data is going to color how we view the next 10 years. But in reality, it needs to be a lot more of that, a lot more than one year to impact a model substantially. It needs to be a decade out of 100 years, probably at least. So, yeah, totally makes sense. And on that, I think we know that risk intelligence requires complex multilayered data. As we've just covered, the context and understanding comes from the interplay of a wide variety of data types from various sources and inputs. Operationally, a human underwriter can't manually process that volume and that volume of information without losing context. So could you talk a little bit more on how the emergence of LLMs and agentic AI tools create a perfect synergy with Geoscope's complex data sets to deliver instant but also actionable intelligence? Yeah, I'd like to talk about that. But before that, right today, the UN, the United nations, they just published a post on LinkedIn. I was this morning, as always, with the phone finger up and down, checking, linking, and they were talking exactly about that, the un, that AI will empower complex climate risk assessments. This is from the un. And another thing that was in that post, actually. But be careful with the black box. That's also another thing. And here I would say also with regards to that post, black box. Well, indeed, on one hand, if you just have a black box risk index, what we are talking about, dashboard, AI can do nothing because the insight is already done. The insight is there, everything is already cooked. Nothing to do, it's done. Yeah, but if you have the complexity and you have the context, AI can do some pretty amazing things. But there is also the factor of explainability. Xai? Yeah, the explainability of the AI, the transparency of the AI and so on and so forth. So I believe that the post of the UN was kind of going like, on, AI can do amazing things, but take care of black box. I would say on both factors, because there are two Types of black box here. Yeah. But on the other hand, on our experience, and I think I mentioned at the beginning with our platform sustex, with our climate data, that we generate kind of climate infrastructure, at the beginning, it was frustrating. Back in early 2025, late 2024, this was like, that's super cool. But I understand nothing. Yeah, it's like I expect a climate risk index like 3 out of 5 or a green or a red. And now we are at the time in which we have to shift and everything is shifting into. Let's just focus on providing the right context to this AI. What are my expectations? And let's try to assemble or get the maximum amount of data. Maybe we can specialize this AI into different field agents and so they can work together. That could be, as we mentioned, the soil expert, the climate expert, but indeed all this data complexity that we were promoting at the very beginning and that it was scary for our clients, which is the worst thing that you could have. It's becoming more and more and more attractive. And that's super good news. I will not say that it was like, oh, I envisioned that back in 2023, I was convinced that it would be that way. And so that's why I went all in. We have 25 terabytes of data, proprietary data, climate data, purely climate data, 25 terabytes, the global scale. It was just our nature. And also it's a big thanks as I am here for the European Space Agency because they kind of believe in us. Like, hey, we have to do things right. And for any risk assessment, for any underwriter, for any physical risk evaluation, it cannot be, it will never be a traffic light. So since the very beginning we had this idea. We also focus a lot in the transparency. We are the only company out there that you can log in into our platform. And one of the things that you can see visually is for each of our models, where we are better or where we are worse. So we show the curacy. It's not that we're talking about the curacy. It's like since the very beginning we were cross validating the accuracy and we're showing that how good or how bad we are, we know that's crucial. And maybe for an average human like me, or maybe I'm even below the average, but this tough to work with, but not for an AI. It can weight, let's say, this accuracy metric, it can adapt this accuracy metric to a specific use case. So all this supplementary data that we were putting, that was scary for the end user or the client for the underwriter, for the consultant or the expert, whoever. At some point it's becoming more and more attractive because now we have the help of the AI and it's what we mentioned. And every time more and more it's not about doing the job, it's about giving the critical thinking. And the critical thinking is, is this traffic light useful? Is this traffic light worth anything? If I have to say it this way, that's part of the critical thinking. Right. Because the AI can do it. The AI can do the job. And as humans we have to add the critical thinking. Or if I get suspects, data or data. Is this data realistic? Maybe you can prompt these questions to the AI or you can cross validate it. Not just to let an automatic workflow or prepare the workflow to check scientific evidence, check reports. I mean we have to say critical thinking, thinking for everybody, for us two, we are very exposed and we are very open to get feedback. Critical thinking, human critical thinking rather than AI reasoning, but human critical thinking. Oh, hey, I've been testing your data. I don't get this, I don't get this other thing. How can I solve this? We are also open sourcing Python code so that people can use systex in a very automated manner. So our goal is also to promote, indeed you can put it in the MCP in the model context protocol. Yeah. Of an AI, but it's also to promote the human interaction with all this data and to promote the critical thinking. So LLMs have a big role and humans too, because they have to. Like our ancient did. Yeah, civilization, like the Greeks did critical thinking, that was something that they were respecting and promoting and it's time to go back to that. So yeah, that's the shared role, shared responsibility that I see here. And that so far it was just, oh, let's do the job, let's just underwrite, let's just take this silo data and maybe this silo data. Let's just see if there is any standard that we can combine it. Oh, that's a standard. Blah, blah, blah. I'll use it. Is it representative of your use case? Because as we said, wildfires are not the same in Canada than in Andalusia. And maybe you're using a standard for Canada and you're applying this risk to Andalucia with a very different climatology soil type. So the part of the critical thinking is very important and I think that AI will empower, will make us better in that because we were so focused on the Task on the job. So that's why also my thoughts and my view on that shift to more optimistic. Yeah, and it's a good point. We talk about AI and LLMs and how we can use them almost in a vacuum, but there's two pieces to that puzzle. There's the AI, the LLM and the output, but then there's the human interaction with it. And as you say, I mean, to really kind of drive progress in that region, there's a reframing of approaches needed. But equally for humans to interact with that in a way where they trust the output, it needs to be explainable. They need to understand the rationale and the reason and the inputs that have led to a certain conclusion, which is, I mean, I totally agree. That highlights why black box type approaches where they're not going to cut it in the coming years. Because we need to understand what's going on under the hood to truly, truly engage with it as humans. And I guess on these new AI capabilities, many large carriers have technical teams that want to build their own AI agents or data integrations entirely in house and from scratch, which is understandable. These teams, they're passionate about this and it's what they enjoy doing. But from your perspective, what are the inherent risks and limitations of that pure build approach? And why should insurers instead be looking to specialist partners for their risk intelligence strategies? So there are risk. Some risks are very notable. Yeah. And there are a few advantages, but I'm just trying to put both in the both sides, advantages and risk. But indeed, we are not alone in the planet. And this is something that it's a reflection that we did as sustex, as Geoscope, as Geoscope, proprietary of Sustex, of this climate data. Because at some point the easy answer is like, oh, everything is shifting to AI and I have the data. Cool. I believe that the data is what the society or what AI requires. And so AI can generate the insight out of data and adapt it. So why as source text? Why not putting an AI agent on my platform? And so people can just type the question that whatever they are interested and get an insight. Why not? This would go against our view of what we mentioned that we need when we talk about risk, when we talk about underwriting risk, climate risk assessment, all of this. We need a holistic view in terms of complexity of the climate data, in terms of socioeconomic data, in terms of health data, in terms of soil data. I mean everything. We need a holistic view. If I now go myself and I put AI into my Platform, I am contradicting myself. And so this could answer at some point the question, if we take this example, this situation, into carriers or into insurers. Insurers, they have their own data, but this data could be skilled, could be skilled due to their market capabilities. So which countries are you located? If you have maybe more male data, female data, and this could skew the analysis that AI could do, then the insurer that is not doing that, that is taking multiple data sources, that is more open and instead of going on the silo mode, doing it also with partners, will have a super big competitive advantage because we'll have more insights, economic insights, climate insights, and these will empower a lot decision making. The other insurer that is taking this silo approach, oh, I want my own AI will always have this limitation that could eventually will lead to errors. The advantage indeed would be maybe, but I'm not that sure because in the end, the cloud is a cloud. You're owner of your data, but in the end, in most cases it's on the cloud anyway. So it's on Azure or Google Cloud. And we all know what this means in the end. Yeah, whatever the contract says. So at some point it's more like there's a lot of disadvantages if you go, oh, I'm going to develop my own AI agents because I want to be owner of the data, and so on. And even if you go partnership or if you go with other stakeholders that can provide you the intelligence, the AI agents or data sets, it's always easy to handle. Let's say this privacy or this serenity that you would be looking for the data. So that's a very big disadvantage. I would say that your vision will be skewed compared to what your neighbor will have. And that's what we saw in sustex. That's why we are partnering with you guys. Because in the end it's like, we understand that, okay, we have a lot of data, we have complex data, but it's not only about this, it's not only about this. It's also about using multiple other data sources. And we just want to be one of the pilots. But there could be like 20 or 40 more pilots in just one decision. We need the maximum flexibility for make it useful. That's how we would really make good decisions. That's what we are convinced of. And this is why we didn't go on fancy dashboards or we didn't go on AI agents inside systex. So we took a more, let's say, democratic approach. Interesting. And I mean, I think one of the key messages in that is that the task of creating an agentic AI ecosystem that functions, it's a huge task and that's one of the drawbacks. But I think creating it is the first step, scaling it is even harder and ensuring that it kind of works consistently and can be maintained is just an astronomical task. So I mean you've painted a picture of all the different moving parts involved and I think that really exemplifies why the purely build approach is perhaps not one to commit to for a carrier because I mean they have other priorities, they have have smaller teams perhaps than some people that focus on this specifically. So yeah, really interesting insight. Thank you. And finally, if an insurer successfully transitions away from generic reporting dashboards and embraces true AI driven climate intelligence, how does that fundamentally change their confidence and capability when evaluating and pricing long term and particularly multi year commercial contracts? That's a very good question. This cannot be, let's say, quantified. The answer, it cannot be quantitative. Even I love to, because it's relative on where you're applying this climate intelligence. It's relative. So I believe that you always have to use the standard procedure, the historical data. So that's one of the things. But on the other hand, this historical Data is not 100% representative, not even because of the climate change, let's call it the climate dynamics. The climate is not static per se. So you need projections or you need models of the earth that are capable to interpret this dynamism. So that's a source of climate intelligence. Going all in with only historical data and following the traditional approaches and modeling return periods with generalized extreme Jeff or wavelength distribution or whatever, just because of the climate dynamism, but also the climate change will be skewed. And using only projections is kind of crazy because it's a projection. In the end, it's a projection of what we expect. I believe that the true climate intelligence lies on the combination of both factors. And also when we are following the standard procedures that use historical data and we are modeling this wavable distribution or this Jeff, generalized extreme model distribution, we are assuming that this way of modeling is the most correct one, which also could be wrong. Could be not the most correct one. Maybe instead of SGF you need another type of distributions or you need another type of feeds and so on. And here is also where an AI could add value. And that's something that I've been playing myself to just for fun, let's say. But what's the best way to model this data, to model this risk, to model this return period, to model this flood, to model these heat waves. There are some standards, but in some cases this could be wrong. So embracing the true climate intelligence involve feeding your AI with the whole context, the past, the future. We also provide, for instance in Systex, the past of our projections, so that it can be comparable. We provide, let's say, a baseline for the past, which is a model, not a big Secret. Reinologies Model Era 5 and then the same projection per se. So there's a comparison point. And if we also provide all this component of the past, because we are saying about this reality small, it's because we believe it should be used too. Now how to use it? I would go on a very holistic approach. Again, taking the past, taking all the future test scenarios, and fitting all of these into AI with the context. The standard, the designer standard, uses this distribution, uses this model. Can you check other models? All of this indeed can be automated as long as you have an agent. So providing this context, this rational, this critical thinking is super important. And then the powers of automation is huge. And that's real climate intelligence. Not using only the climate projection, not using only the pass, using both being aware of the context and providing it to these AI agents in Pluto, because then you will have the statistination, the climate specialist and whatever. And from here you can take insights that I'm convinced that in 80 to 90% of the cases will be much accurate and much more representative. Could be the same return period, but when you were using only historical data, maybe you had that storm of 1 on 100 return period involve, I don't know, wind gust of 32 meters per second or 35 meters per second. Once you combine both, the insight that you would get is like, no, it involves 38, with this uncertainty from 36 to 40, assuming the past, the future, weighting the past and weighting the future, sorry, based on the past data and so on and so forth, this is what the AI can do. Now that's magic, that's super cool. Because for an insurer, in the end you have the uncertainty branches, you have the expected. And so you can model, you can put stop losses and whatever. So you can do some very cool things out of this. So that's the true climate intelligence. And that's where, and especially if later on you can put economic, socioeconomic data into this. So which kind of assets are we talking about? How they were designed? Are they ready to support the worst case of 40 meters per second? Or they are not because we have this, let's say threshold of uncertainty now. So that's a true climate intelligence. And again it's this super holistic view is this super way of providing the context, critical thinking. And from here the power of automation is huge. And indeed always checking everything like explainable AI that you were talking about. That's super crucial. I believe so in terms of LLM or AI agents to be able to understand what flow was following. So AI, Asian or my dear friend, tell me, which distribution were you fitting the wind data? Where did you get the socioeconomic data that you can follow that? Because you'll know at some point maybe it did something very, very wrong. Like wind should have always a wavel distribution and it's doing a Gaussian. So that's super wrong. So you need also this factor of explainability and this factor of control that would help you fine tune also the context. Yeah, interesting. I guess what we're trying to do is predict the future and the only way to do that is with as much data as you can possibly get your hands on that's relevant. As you say, taking historical data, future modeling, but ensuring it's over a large enough timescale to really kind of be reliable and to not be skewed by anomalies and to accurately kind of predict, particularly for underwriters, what's going to happen during the, I guess the timescale of their contracts. It's quite easy to see why that's so essential to them. And we're taking really large data sets and trying to understand what they're going to do to a 1, 2, 3 year period of time. It really is a fine, I mean, you said it's a science and it certainly is, but I think it's a fine art as well to try and achieve that. So thank you for those insights, they've been fantastic. And there's one more thing I do want to ask you today, please. I'm loving it. So Azim, I. If you could change one thing about the insurance industry, what would that one thing be and why? So about the insurance industry, I believe there is a need of openness overall. So it's a crucial industry for our society. If something is not insurable, no bank will finance this something if there is nobody ready to insure this. So it will be not develop. So if a hospital cannot venture, you have a big problem. So I'm trying not to go in my philosophical mode, but I'm trying to think about what's important for the insurance industry or what I would ask because it's a very crucial industry for our society. Honestly speaking, openness means like being keen on accepting new data, being keen on validating what you get. So openness from a critical standpoint, this is something that it's super important. And indeed, another thing I would say is stop relying on traffic lights and nutritional score indexes. Like, even for esg, they are not worth it. It's an oversimplification that it's not worth it. In the end, we have plenty of examples open on our website on how to analyze a risk code to do that, because we don't want this to happen. I mean, even happy to give it away for free. But please stop relying on just let's say this traffic line, this oversimplificated indices, free scores that in the end they are not worth it. Yeah, no, I agreed. I think the openness you described is one of the key factors of something we discussed earlier, which is that's one of the approaches and the, I guess, qualities that those that lead are going to need when it comes to insurance. Those that are open to new methods and to working with partners to validate data and to embracing complexity and new technologies, they're the ones that are going to pull ahead. And I totally agree with you there. And I think to end on a very relevant note, I think abandoning the traffic light system is another key aspect for insurance. So, Johan, thank you for joining us today. I've really enjoyed our conversation. It's been super insightful. A pleasure, actually. I mean, we have to do it more. It's been lovely. Thanks a lot for having me here and super happy about the collaboration that we're moving ahead. Absolutely. Me too. So, until the next one and all the best. Thank you, Johan. Thanks. Have a nice day. Cheers. Making Risk Flow is brought to you by Zytora. If you enjoy this podcast, consider subscribing to Making Risk Flow in Apple Podcasts, Spotify or wherever you get your podcast. So you never miss an episode. To find out more about Saitora, visit siteora.com thanks for joining me. See you next time.